Multi-Model Splicer
The Splicer sends one prompt to several models in parallel, then judges their answers for consensus. Instead of trusting a single model, you get an agreement score, a synthesized answer, and a list of where the models disagreed.
Use it when the answer matters: architecture decisions, legal/medical-adjacent questions, fact-heavy research, anything where one model hallucinating would hurt.
How to run a splice
- In chat, open the Splicer panel
- Pick your models — any model from the full catalog, up to your tier's limit
- Send your prompt once
The prompt fans out to every selected model simultaneously. When the responses land, a judge model compares them and the consensus card renders:
- Agreement % — how much the models' answers align
- Synthesized answer — a single best answer combining the cohort
- Divergences — specific points where models disagreed, so you know exactly what to double-check
Splicer runs are persisted — reload the conversation and the full consensus card comes back, including costs and savings.
What it costs
You pay the actual usage cost of each model in the fan-out (plus the judge). Cost comes from real per-token catalog pricing, summed across the cohort — a 4-model splice on cheap models can cost less than one premium-model reply.
Compression is on by default. Before fan-out, your prompt runs through the compression engine once, and the optimized prompt is what fans out — so the token savings multiply by the number of models. The consensus card shows the result, e.g. "saved $0.004 · 38% via compression." Long prompts routinely save ~40% of input tokens. You can opt out per-request with splicer.compress: false.
Tier limits
| Tier | Max models per splice | Daily Splicer cap |
|---|---|---|
| Free | — (not available) | — |
| Pro | 4 | $5.00 / day |
| Teams | 6 | $25.00 / day |
| Enterprise | unlimited | none |
| API | unlimited | usage-based |
The daily cap counts actual splice spend and resets at midnight UTC. If a splice would exceed the cap you'll get a clear error rather than a partial run.
Developer API
Splices run through the chat interface, but two read endpoints exist:
| Method | Path | Purpose |
|---|---|---|
| GET | /v1/splicer/registry | Models available to splice for your tier |
| GET | /v1/splicer/runs/{id} | Full run detail — per-model responses, consensus, costs, compression savings |
How it relates to Consensus Guard
The platform continuously samples a share of routed traffic in the background, fanning it to alternate models and scoring quality (that's Consensus Guard — it's how the auto-router learns which models are actually good at which domains). The Splicer is the same machinery put directly in your hands: you choose when a prompt deserves the multi-model treatment.
Where to next
- Auto-Router: Tiers, Domains, and Providers — how single-model routing works
- Picking a Model — manual model selection
- Wallet and Funds — funding splice runs